Contrastive Learning for Cold Start Recommendation with Adaptive Feature Fusion

2025-02-05

AI Summary

This work addresses representation degradation and performance decline in recommender systems under cold-start scenarios, where sparse user–item interactions hinder effective learning. To tackle this, we propose a novel framework integrating contrastive learning with adaptive multimodal feature fusion. Methodologically: (1) we jointly model adaptive feature weighting and contrastive learning to dynamically integrate user attributes, item metadata, and contextual features—a first in the literature; (2) we design a sparse-interaction-aware positive/negative sampling strategy to enhance representation robustness and generalization. Experiments on MovieLens-1M demonstrate significant improvements over baselines—including MF, DeepFM, and AutoRec—with HR@10 and NDCG@10 increasing by up to 12.7% on the cold-start subset. Ablation studies confirm both the effectiveness and complementary nature of each component.

Technology Category

• Data Mining & Knowledge Management: Recommender Systems
• Machine Learning: Representation Learning
• Computer Vision: Representation Learning for Vision

Application Category

• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs

Abstract

This paper proposes a cold start recommendation model that integrates contrastive learning, aiming to solve the problem of performance degradation of recommendation systems in cold start scenarios due to the scarcity of user and item interaction data. The model dynamically adjusts the weights of key features through an adaptive feature selection module and effectively integrates user attributes, item meta-information, and contextual features by combining a multimodal feature fusion mechanism, thereby improving recommendation performance. In addition, the model introduces a contrastive learning mechanism to enhance the robustness and generalization ability of feature representation by constructing positive and negative sample pairs. Experiments are conducted on the MovieLens-1M dataset. The results show that the proposed model significantly outperforms mainstream recommendation methods such as Matrix Factorization, LightGBM, DeepFM, and AutoRec in terms of HR, NDCG, MRR, and Recall, especially in cold start scenarios. Ablation experiments further verify the key role of each module in improving model performance, and the learning rate sensitivity analysis shows that a moderate learning rate is crucial to the optimization effect of the model. This study not only provides a new solution to the cold start problem but also provides an important reference for the application of contrastive learning in recommendation systems. In the future, this model is expected to play a role in a wider range of scenarios, such as real-time recommendation and cross-domain recommendation.

Problem

• Cold start recommendation performance
• Scarcity of interaction data
• Feature fusion and contrastive learning

Innovation

• Contrastive learning enhances feature robustness
• Adaptive feature fusion optimizes key feature weights
• Multimodal feature integration improves recommendation accuracy